• Each problem must be self-contained in a file, named properly, and must compile using the command mpicc filename.c
• On this homework you may not use any of the standard default MPI collective operations, you must only use MPI Send and MPI Recv.

Problems

Level:Basic

1. Create a function my scatter with declaration

void my_scatter(int* sendbuf, int sendcount, int* recvbuf, int recvcount, int root)

that re-creates some of the functionality of MPI Scatter, assuming that they send type is MPI INT, and that the communicator is MPI COMM WORLD. Note that sendcount and recvcount might not match. After all, if rank j is not the root, then rank j would reasonably say that its send count is 0, while its receive count is non-zero. Your code must follow the simple algorithm of repeatedly sending from the root to every non-root process; see Figure 1. The root should not send to itself, but should rather copy the data into the correct buffer. You may not use MPI Scatter. Call the file my scatter.c

2. Create a function my broadcast with declaration

void my_broadcast(int* buffer, int count)

that re-creates some of the functionality of MPI Bcast, assuming that they send type is MPI INT, that the root is 0, and that the communicator is MPI COMM WORLD. Call the file my broadcast.c. Your code may assume that the number of processors on the communicator is a power of 2. Your code should follow the fan-out algorithm from class see Slides: Collective Operations, slide 22 on Moodle. You may not use MPI Bcast. Call the file my broadcast.c.

3. Create a function my allgather with declaration

Figure 1: A simple scatter algorithm

void my_allgather(int* sendbuf, int sendcount, int* recvbuf)

that reproduces some of the functionality of MPI Allgather, assuming that the send type is MPI INT, that send count and receive count match, and that the communicator is

MPI COMM WORLD. Again, you may only use MPI Send and MPI Recv for communication. In this code, each process should do k 1 sends and k 1 receives, where k is the number of processors engaged see slide 28 in Slides: Collective Operations on Moodle.

Level: Advanced

1. Create a function my gather with declaration

void my_gather(int* recvbuf, int recvcount, int* sendbuf, int sendcount, int root)

that improves the my gather function we wrote in class so that it does not assume that the number of processors is a power of 2, and it does not assume that the root is rank 0. Your algorithm should follow the fan-in algorithm discussed in class, but treating root as rank 0, root+1 as rank 1, and so on. Name the file my gather.c

2. Create a function my allplus with declaration

int my_allplus(int* sendbuf, int* recvbuf, int count)

that puts the sum of all the data held in each processors sendbuf array on every process.

For example, if there are four processes p₀,p₁,p₂,p₃ and p_i holds data d_i,0,d_i,1,d_i,2 then when the processes complete the call to my allplus each processor should find

d0,0 + d0,1 + d0,2 + d1,0 + d1,1 + d1,2 + d2,0 + d2,1 + d2,2 + d3,0 + d3,1 + d3,2 available in recvbuf[0].

No processor should do more than dlogpe receives and dlogpe sends, and no send or receive should be longer than one integer.

Note: This is not (quite) the algorithm that we sketched out in class!

Errata

What does the testing code do, and how to interpret the results

The testing code compiles your function (my all gather, my scatter or whatever), and compiles them along with a main function that I have written. It calls your function with a number of different inputs, and makes sure that your code is behaving correctly. The command that Im using to compile your code is shown first. For example:

mpicc std=c99 -g -lm -Wl,-wrap,MPI_Recv -Wl,-wrap,MPI_Send

-Wl,-wrap,main my_alltoall.c test_helpers.c my_alltoall_test_driver.c mpi_send_wrapper.c -o my_alltoall_exe

You can see that Im calling mpicc to compile your code into a bunch of code that Ive written. You can compile your code exactly the same way I am, if you want. Of course, then you would be using my main rather than the one you wrote.

You can see the arguments that my main is giving your functions in each test case. For example, here is the output of one test case for my alltoall

Using the command: mpirun -np 2 ./my_alltoall_exe len=10 Rank 1 calls my_alltoall with: sendbuff:[1,346,381,486,301] sendcount:5 recvbuff:[-1939952792,32728,26006624,0,26004688,0,25434272,0,1,0] recvcount:5

Rank 0 calls my_alltoall with: sendbuff:[0,243,272,259,420] sendcount:5 recvbuff:[-222065816,32616,13234448,0,13233520,0,12691280,0,1,0] recvcount:5

Rank 0 finishes my_alltoall with:

recv_buff: [0,243,272,259,420,1,346,381,486,301]

Rank 1 finishes my_alltoall with: recv_buff: [0,243,272,259,420,1,346,381,486,301]

We can see that rank 0 has an initial sendbuf of [1,346,381,486,301] and rank 1 has an initial sendbuf of [0,243,272,259,420]. The code completes, and after calling my alltoall both ranks have [0,243,272,259,420,1,346,381,486,301] in their recv buff.

Once the executable runs, my code examines some logs to make sure that your code executed in a reasonable fashion (the correct number of sends, receives, etc).

If you wish to debug a certain test case, you can just hard-code in the inputs into your own main, and call your function with the same inputs that my code is using.

Using the standard code

I have provided some standard implementations of the assigned problems. In order to use them, you will need to replace your calls to foo with foo standard. For example, if you want to see how my my alltoall function behaves, just replace your calls to my alltoall with calls to my alltoall standard.

When you compile your code, you will need to add in the object files provided in the obj directory.

For example, you would switch from compiling my alltoall like this:

mpicc my_alltoall.c -o my_alltoall

to like this:

mpicc my_alltoall.c ./objs/my_alltoall_standard.x86_64.o -o my_alltoall

The compiler will warn you about implicit declaration when you compile, but the overall compilation will succeed.